The problem: Left to its own habits, an autonomous or long-running session drifts back to the expensive model one convenient call at a time. Nothing forces the drift, it just happens, and by the time the weekly quota is gone you didn’t get nearly as much done as the quota should have bought.
qq-go-afk-lean is a standing posture, not a one-off routing decision. Every unit of work defaults to the cheapest capable model: Groq or Cerebras first (free), Ollama for anything private, DeepSeek Flash as the paid anchor that essentially never rate-limits. Claude is reserved for exactly four things: final synthesis of what the cheap models produced, voice and copy (the one place cheap models genuinely damage quality instead of just costing more), irreversible or architectural judgment, and deep reasoning an external model has already demonstrably failed at, never preemptively. On a 429 or a timeout, the rule is always “try a different external model,” never “fall back to Claude to just get it done,” since that defeats the entire posture. It composes with pace, not against it: a bounded one-task-at-a-time run and a wide fan-out run are both still lean as long as the workers doing the actual work are external and Claude only judges the result. This is the frugal end of the spectrum: protect the token budget above almost everything else. For the unbounded end, use qq-go-afk-ham. For the adaptive middle, use qq-go-afk-smart.
Part of Operator Skills, a collection of skills for running a real Claude Code setup.